시각 세계는 어수선하며, 견고한 비전-언어 추론을 위해서는 무관한 정보를 무시해야 한다. 그러나 비전-언어 모델(VLM)은 방해 요소가 존재하는 상태에서 대상 이미지에 관한 질문에 답할 때 상당한 정확도 손실을 겪을 수 있다. 기존 완화 방법은 관련 이미지를 선택하거나 이미지 간 정보 교환을 제한하는 데 초점을 둔다. 방해 이미지를 포함하도록 변형한 다섯 가지 인식 벤치마크에서 공개된 7-8B 규모 VLM 세 종을 체계적으로 분석한 결과, 모델이 관련 이미지를 올바르게 식별하는 경우에도 방해 요소가 정확도를 떨어뜨림을 보이며, 이는 선택만으로는 충분하지 않음을 시사한다. 이러한 간섭의 메커니즘을 이해하기 위해 Qwen2.5-VL에서 어텐션 분석과 표적 녹아웃을 결합하였고, 중간 계층의 방해 요소-텍스트 연결이 무관한 시각 정보가 추론에 영향을 미치는 경로임을 확인하였다. 또한 방해 요소가 있을 때와 없을 때의 답변 분포가 일관되도록 유도하는 답변 일관성 정규화를 도입한다. 이 목적 함수는 어텐션을 직접 제약하지 않으면서도 질문 토큰에서 방해 이미지로 향하는 중간 계층 어텐션을 감소시키며, 이는 분석에서 제시된 경로와 일치한다. 방해 요소가 다섯 개일 때, 제안 방법은 다섯 벤치마크 전반에서 Qwen2.5-VL 기준선 대비 순환 정확도를 평균 16 퍼센트포인트 향상시키면서 단일 이미지 성능을 유지한다.
The ability to interpret and explain complex information from visual data in charts is crucial for data-driven decision-making. In this work, we address the challenge of providing explanations alongside answering questions about chart images. We present ChartQA-X, a comprehensive dataset comprising various chart types with 28,299 contextually relevant questions, answers, and detailed explanations. These explanations are generated by prompting six different models and selecting the best responses based on metrics such as faithfulness, informativeness, coherence, and perplexity. Our experiments show that models fine-tuned on our dataset for explanation generation achieve superior performance across various metrics and demonstrate improved accuracy in question-answering tasks on new datasets. By integrating answers with explanatory narratives, our approach enhances the ability of intelligent agents to convey complex information effectively, improve user understanding, and foster trust in the generated responses.
The Visual Dialogue task requires an agent to engage in a conversation about an image with a human. It represents an extension of the Visual Question Answering task in that the agent needs to answer a question about an image, but it needs to do so in light of the previous dialogue that has taken place. The key challenge in Visual Dialogue is thus maintaining a consistent, and natural dialogue while continuing to answer questions correctly. We present a novel approach that combines Reinforcement Learning and Generative Adversarial Networks (GANs) to generate more human-like responses to questions. The GAN helps overcome the relative paucity of training data, and the tendency of the typical MLE-based approach to generate overly terse answers. Critically, the GAN is tightly integrated into the attention mechanism that generates human-interpretable reasons for each answer. This means that the discriminative model of the GAN has the task of assessing whether a candidate answer is generated by a human or not, given the provided reason. This is significant because it drives the generative model to produce high quality answers that are well supported by the associated reasoning. The method also generates the state-of-the-art results on the primary benchmark.
We propose a technique for making Convolutional Neural Network (CNN)-based models more transparent by visualizing input regions that are 'important' for predictions -- or visual explanations. Our approach, called Gradient-weighted Class Activation Mapping (Grad-CAM), uses class-specific gradient information to localize important regions. These localizations are combined with existing pixel-space visualizations to create a novel high-resolution and class-discriminative visualization called Guided Grad-CAM. These methods help better understand CNN-based models, including image captioning and visual question answering (VQA) models. We evaluate our visual explanations by measuring their ability to discriminate between classes, to inspire trust in humans, and their correlation with occlusion maps. Grad-CAM provides a new way to understand CNN-based models. We have released code, an online demo hosted on CloudCV, and a full version of this extended abstract.
What makes an image appear realistic? In this work, we are answering this question from a data-driven perspective by learning the perception of visual realism directly from large amounts of data. In particular, we train a Convolutional Neural Network (CNN) model that distinguishes natural photographs from automatically generated composite images. The model learns to predict visual realism of a scene in terms of color, lighting and texture compatibility, without any human annotations pertaining to it. Our model outperforms previous works that rely on hand-crafted heuristics, for the task of classifying realistic vs. unrealistic photos. Furthermore, we apply our learned model to compute optimal parameters of a compositing method, to maximize the visual realism score predicted by our CNN model. We demonstrate its advantage against existing methods via a human perception study.
Recently, reinforcement learning (RL) has been shown to greatly enhance the reasoning capabilities of large language models (LLMs), and RL-based approaches have been progressively applied to visual multimodal tasks. However, the audio modality has largely been overlooked in these developments. Thus, we conduct a series of RL explorations in audio understanding and reasoning, specifically focusing on the audio question answering (AQA) task. We leverage the group relative policy optimization (GRPO) algorithm to Qwen2-Audio-7B-Instruct, and our experiments demonstrated state-of-the-art performance on the MMAU Test-mini benchmark, achieving an accuracy rate of 64.5%. The main findings in this technical report are as follows: 1) The GRPO algorithm can be effectively applied to large audio language models (LALMs), even when the model has only 8.2B parameters; 2) With only 38k post-training samples, RL significantly outperforms supervised fine-tuning (SFT), indicating that RL-based approaches can be effective without large datasets; 3) The explicit reasoning process has not shown significant benefits for AQA tasks, and how to efficiently utilize deep thinking remains an open question for further research; 4) LALMs still lag far behind humans auditory-language reasoning, suggesting that the RL-based approaches warrant further exploration. Our project is available at https://github.com/xiaomi-research/r1-aqa and https://huggingface.co/mispeech/r1-aqa.
Vision-language (VL) understanding tasks evaluate models' comprehension of complex visual scenes through multiple-choice questions. However, we have identified two dataset biases that models can exploit as shortcuts to resolve various VL tasks correctly without proper understanding. The first type of dataset bias is \emph{Unbalanced Matching} bias, where the correct answer overlaps the question and image more than the incorrect answers. The second type of dataset bias is \emph{Distractor Similarity} bias, where incorrect answers are overly dissimilar to the correct answer but significantly similar to other incorrect answers within the same sample. To address these dataset biases, we first propose Adversarial Data Synthesis (ADS) to generate synthetic training and debiased evaluation data. We then introduce Intra-sample Counterfactual Training (ICT) to assist models in utilizing the synthesized training data, particularly the counterfactual data, via focusing on intra-sample differentiation. Extensive experiments demonstrate the effectiveness of ADS and ICT in consistently improving model performance across different benchmarks, even in domain-shifted scenarios.
In the construction sector, workers often endure prolonged periods of high-intensity physical work and prolonged use of tools, resulting in injuries and illnesses primarily linked to postural ergonomic risks, a longstanding predominant health concern. To mitigate these risks, researchers have applied various technological methods to identify the ergonomic risks that construction workers face. However, traditional ergonomic risk assessment (ERA) techniques do not offer interactive feedback. The rapidly developing vision-language models (VLMs), capable of generating textual descriptions or answering questions about ergonomic risks based on image inputs, have not yet received widespread attention. This research introduces an interactive visual query system tailored to assess the postural ergonomic risks of construction workers. The system's capabilities include visual question answering (VQA), which responds to visual queries regarding workers' exposure to postural ergonomic risks, and image captioning (IC), which generates textual descriptions of these risks from images. Additionally, this study proposes a dataset designed for training and testing such methodologies. Systematic testing indicates that the VQA functionality delivers an accuracy of 96.5%. Moreover, evaluations using nine metrics for IC and assessments from human experts indicate that the proposed approach surpasses the performance of a method using the same architecture trained solely on generic datasets. This study sets a new direction for future developments in interactive ERA using generative artificial intelligence (AI) technologies.
We present VARGPT, a novel multimodal large language model (MLLM) that unifies visual understanding and generation within a single autoregressive framework. VARGPT employs a next-token prediction paradigm for visual understanding and a next-scale prediction paradigm for visual autoregressive generation. VARGPT innovatively extends the LLaVA architecture, achieving efficient scale-wise autoregressive visual generation within MLLMs while seamlessly accommodating mixed-modal input and output within a single model framework. Our VARGPT undergoes a three-stage unified training process on specially curated datasets, comprising a pre-training phase and two mixed visual instruction-tuning phases. The unified training strategy are designed to achieve alignment between visual and textual features, enhance instruction following for both understanding and generation, and improve visual generation quality, respectively. Despite its LLAVA-based architecture for multimodel understanding, VARGPT significantly outperforms LLaVA-1.5 across various vision-centric benchmarks, such as visual question-answering and reasoning tasks. Notably, VARGPT naturally supports capabilities in autoregressive visual generation and instruction-to-image synthesis, showcasing its versatility in both visual understanding and generation tasks. Project page is at: \url{https://vargpt-1.github.io/}
Deep models are the defacto standard in visual decision problems due to their impressive performance on a wide array of visual tasks. On the other hand, their opaqueness has led to a surge of interest in explainable systems. In this work, we emphasize the importance of model explanation in various forms such as visual pointing and textual justification. The lack of data with justification annotations is one of the bottlenecks of generating multimodal explanations. Thus, we propose two large-scale datasets with annotations that visually and textually justify a classification decision for various activities, i.e. ACT-X, and for question answering, i.e. VQA-X. We also introduce a multimodal methodology for generating visual and textual explanations simultaneously. We quantitatively show that training with the textual explanations not only yields better textual justification models, but also models that better localize the evidence that support their decision.
Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available (https://reasoningpatterns.github.io). We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.
In this paper, we present the mQA model, which is able to answer questions about the content of an image. The answer can be a sentence, a phrase or a single word. Our model contains four components: a Long Short-Term Memory (LSTM) to extract the question representation, a Convolutional Neural Network (CNN) to extract the visual representation, an LSTM for storing the linguistic context in an answer, and a fusing component to combine the information from the first three components and generate the answer. We construct a Freestyle Multilingual Image Question Answering (FM-IQA) dataset to train and evaluate our mQA model. It contains over 150,000 images and 310,000 freestyle Chinese question-answer pairs and their English translations. The quality of the generated answers of our mQA model on this dataset is evaluated by human judges through a Turing Test. Specifically, we mix the answers provided by humans and our model. The human judges need to distinguish our model from the human. They will also provide a score (i.e. 0, 1, 2, the larger the better) indicating the quality of the answer. We propose strategies to monitor the quality of this evaluation process. The experiments show that in 64.7% of cases, the human judges cannot distinguish our model from humans. The average score is 1.454 (1.918 for human). The details of this work, including the FM-IQA dataset, can be found on the project page: http://idl.baidu.com/FM-IQA.html
With the emergence of large language models (LLMs) and vision foundation models, how to combine the intelligence and capacity of these open-sourced or API-available models to achieve open-world visual perception remains an open question. In this paper, we introduce VisionGPT to consolidate and automate the integration of state-of-the-art foundation models, thereby facilitating vision-language understanding and the development of vision-oriented AI. VisionGPT builds upon a generalized multimodal framework that distinguishes itself through three key features: (1) utilizing LLMs (e.g., LLaMA-2) as the pivot to break down users' requests into detailed action proposals to call suitable foundation models; (2) integrating multi-source outputs from foundation models automatically and generating comprehensive responses for users; (3) adaptable to a wide range of applications such as text-conditioned image understanding/generation/editing and visual question answering. This paper outlines the architecture and capabilities of VisionGPT, demonstrating its potential to revolutionize the field of computer vision through enhanced efficiency, versatility, and generalization, and performance. Our code and models will be made publicly available. Keywords: VisionGPT, Open-world visual perception, Vision-language understanding, Large language model, and Foundation model
Today's VQA models still tend to capture superficial linguistic correlations in the training set and fail to generalize to the test set with different QA distributions. To reduce these language biases, recent VQA works introduce an auxiliary question-only model to regularize the training of targeted VQA model, and achieve dominating performance on diagnostic benchmarks for out-of-distribution testing. However, due to complex model design, these ensemble-based methods are unable to equip themselves with two indispensable characteristics of an ideal VQA model: 1) Visual-explainable: The model should rely on the right visual regions when making decisions. 2) Question-sensitive: The model should be sensitive to the linguistic variations in questions. To this end, we propose a novel model-agnostic Counterfactual Samples Synthesizing and Training (CSST) strategy. After training with CSST, VQA models are forced to focus on all critical objects and words, which significantly improves both visual-explainable and question-sensitive abilities. Specifically, CSST is composed of two parts: Counterfactual Samples Synthesizing (CSS) and Counterfactual Samples Training (CST). CSS generates counterfactual samples by carefully masking critical objects in images or words in questions and assigning pseudo ground-truth answers. CST not only trains the VQA models with both complementary samples to predict respective ground-truth answers, but also urges the VQA models to further distinguish the original samples and superficially similar counterfactual ones. To facilitate the CST training, we propose two variants of supervised contrastive loss for VQA, and design an effective positive and negative sample selection mechanism based on CSS. Extensive experiments have shown the effectiveness of CSST. Particularly, by building on top of model LMH+SAR, we achieve record-breaking performance on all OOD benchmarks.
대규모 비전-언어 모델(LVLM)은 시각적 비그라운딩(visual ungroundedness) 문제를 겪는다. 이미지가 예측에 전혀 기여하지 않더라도, 순전히 언어 사전분포(prior)만으로 유창하고 자신감 있으며 심지어 정답인 응답을 만들 수 있다. 기존 신뢰도 추정 방법은 일반 추론 하에서 모델 동작만 관찰하며, 예측이 이미지에 의해 형성되었는지 아니면 텍스트만으로 형성되었는지 판별할 메커니즘이 없어 이를 탐지할 수 없다. 우리는 모델에 구애받지 않는 신뢰도 추정 프레임워크인 BICR(Blind-Image Contrastive Ranking)을 제안한다. LVLM은 완전히 고정(frozen)된 상태로 두고, 대신 샘플마다 은닉 상태(hidden state)를 두 번 추출한다—한 번은 실제 이미지-질문 쌍으로, 한 번은 질문은 고정한 채 이미지를 검게 가린(blacked out) 경우로. 이어서 실제 이미지 은닉 상태 위에서 작은 별도 프로be를 학습하며, 검게 가린 뷰에서 더 높은 신뢰도에 페널티를 주는 순위 손실(ranking loss)로 정규화하여 시각적 그라운딩을 신뢰성의 신호로 취급하도록 한다. 추론 시에는 실제 이미지에 대한 단일 패스만 사용하므로, BICR은 표준 단일 뷰 프로be 대비 추가 비용이 없다. 시각 질의응답, 객체 환각(hallucination) 탐지, 의료 영상, 금융 문서 이해를 아우르는 벤치마크에서 5개의 최신 LVLM과 7개 기준선에 대해 평가한 결과, BICR은 교정(calibration)과 변별(discrimination) 모두에서 교차 LVLM 평균 최고 성능을 동시에 달성하며, 클러스터 인식(cluster-aware) 분석에도 견고한 통계적으로 유의한 변별 향상을 보이고, 가장 강한 probing 기준선 대비 4–18배 적은 파라미터를 사용한다.
We introduce EgoLife, a project to develop an egocentric life assistant that accompanies and enhances personal efficiency through AI-powered wearable glasses. To lay the foundation for this assistant, we conducted a comprehensive data collection study where six participants lived together for one week, continuously recording their daily activities - including discussions, shopping, cooking, socializing, and entertainment - using AI glasses for multimodal egocentric video capture, along with synchronized third-person-view video references. This effort resulted in the EgoLife Dataset, a comprehensive 300-hour egocentric, interpersonal, multiview, and multimodal daily life dataset with intensive annotation. Leveraging this dataset, we introduce EgoLifeQA, a suite of long-context, life-oriented question-answering tasks designed to provide meaningful assistance in daily life by addressing practical questions such as recalling past relevant events, monitoring health habits, and offering personalized recommendations. To address the key technical challenges of (1) developing robust visual-audio models for egocentric data, (2) enabling identity recognition, and (3) facilitating long-context question answering over extensive temporal information, we introduce EgoButler, an integrated system comprising EgoGPT and EgoRAG. EgoGPT is an omni-modal model trained on egocentric datasets, achieving state-of-the-art performance on egocentric video understanding. EgoRAG is a retrieval-based component that supports answering ultra-long-context questions. Our experimental studies verify their working mechanisms and reveal critical factors and bottlenecks, guiding future improvements. By releasing our datasets, models, and benchmarks, we aim to stimulate further research in egocentric AI assistants.
Although Visual-Language Models (VLMs) have shown impressive capabilities in tasks like visual question answering and image captioning, they still struggle with hallucinations. Analysis of attention distribution in these models shows that VLMs tend to processing textual tokens rather than visual tokens. This imbalance of attention distribution causes VLMs to favor textual knowledge in the case of multimodal knowledge conflicts, resulting in differences from the image information. In this paper, we propose Re-Balancing Contrastive Decoding (RBD) method, which employs textual and visual branches to recalibrate attention distribution in VLMs. Specifically, the textual branch injects image noise to stimulate the model's dependency on text, thereby reducing textual bias. Concurrently, the visual branch focuses on the selection of significant tokens, refining the attention mechanism to highlight the primary subject. This dual-branch strategy enables the RBD method to diminish textual bias while enhancing visual information. Experimental results demonstrate that our method, RBD, outperforms the existing methods by the CHAIR and POPE metrics, mitigate hallucinations without reducing the model's general capabilities.
Visual Question Answering (VQA) is a fundamental task in computer vision and natural language process fields. Although the ``pre-training & finetuning'' learning paradigm significantly improves the VQA performance, the adversarial robustness of such a learning paradigm has not been explored. In this paper, we delve into a new problem: using a pre-trained multimodal source model to create adversarial image-text pairs and then transferring them to attack the target VQA models. Correspondingly, we propose a novel VQAttack model, which can iteratively generate both image and text perturbations with the designed modules: the large language model (LLM)-enhanced image attack and the cross-modal joint attack module. At each iteration, the LLM-enhanced image attack module first optimizes the latent representation-based loss to generate feature-level image perturbations. Then it incorporates an LLM to further enhance the image perturbations by optimizing the designed masked answer anti-recovery loss. The cross-modal joint attack module will be triggered at a specific iteration, which updates the image and text perturbations sequentially. Notably, the text perturbation updates are based on both the learned gradients in the word embedding space and word synonym-based substitution. Experimental results on two VQA datasets with five validated models demonstrate the effectiveness of the proposed VQAttack in the transferable attack setting, compared with state-of-the-art baselines. This work reveals a significant blind spot in the ``pre-training & fine-tuning'' paradigm on VQA tasks. Source codes will be released.
Dialog is an effective way to exchange information, but subtle details and nuances are extremely important. While significant progress has paved a path to address visual dialog with algorithms, details and nuances remain a challenge. Attention mechanisms have demonstrated compelling results to extract details in visual question answering and also provide a convincing framework for visual dialog due to their interpretability and effectiveness. However, the many data utilities that accompany visual dialog challenge existing attention techniques. We address this issue and develop a general attention mechanism for visual dialog which operates on any number of data utilities. To this end, we design a factor graph based attention mechanism which combines any number of utility representations. We illustrate the applicability of the proposed approach on the challenging and recently introduced VisDial datasets, outperforming recent state-of-the-art methods by 1.1% for VisDial0.9 and by 2% for VisDial1.0 on MRR. Our ensemble model improved the MRR score on VisDial1.0 by more than 6%.
We propose Dual Attention Networks (DANs) which jointly leverage visual and textual attention mechanisms to capture fine-grained interplay between vision and language. DANs attend to specific regions in images and words in text through multiple steps and gather essential information from both modalities. Based on this framework, we introduce two types of DANs for multimodal reasoning and matching, respectively. The reasoning model allows visual and textual attentions to steer each other during collaborative inference, which is useful for tasks such as Visual Question Answering (VQA). In addition, the matching model exploits the two attention mechanisms to estimate the similarity between images and sentences by focusing on their shared semantics. Our extensive experiments validate the effectiveness of DANs in combining vision and language, achieving the state-of-the-art performance on public benchmarks for VQA and image-text matching.